tum-pbs/PhiFlow

A differentiable PDE solving framework for machine learning

What it solves

PhiFlow is a simulation toolkit designed to bridge the gap between physics simulations and machine learning. It allows users to create simulations that are fully differentiable, meaning they can be integrated directly into neural network training loops to optimize physics-based parameters or learn to control physical systems.

How it works

The toolkit is written in Python and integrates tightly with major ML frameworks like PyTorch, Jax, and TensorFlow. By leveraging the automatic differentiation capabilities of these backends, PhiFlow can compute gradients through the simulation process. It provides built-in PDE (Partial Differential Equation) operations, a high-level linear equation solver, and support for various simulation methods including grid-based fluids, particles (SPH, FLIP), and meshes.

Who it’s for

Researchers and developers working on physics-based deep learning, optimization of physical systems, and differentiable physics simulations.

Highlights

  • Multi-backend support: The same simulation code can run on NumPy, PyTorch, Jax, or TensorFlow, enabling easy switching between CPU and GPU execution.
  • Differentiable Simulations: Fully integrates with ML frameworks to allow end-to-end training of models involving physics.
  • Versatile Simulation Types: Supports grid-based fluids, particle simulations, and mesh-based Finite Volume Methods (FVM).
  • Interactive Web UI: Includes a web interface for live visualizations and real-time control of simulations or network training.
  • Dimension Independent: Simulation code is reusable across different dimensions (e.g., 2D vs 3D) without modification.

Related

  • Project
    google/brax

    A fast, fully differentiable physics engine written in JAX for large-scale rigid body simulation and reinforcement learning training.

  • Project
    tumaer/JAXFLUIDS

    A fully-differentiable CFD solver for 3D compressible flows, enabling the integration of machine learning and numerical fluid dynamics through automatic differentiation.

  • Project
    bayesflow-org/bayesflow

    A Python library for efficient amortized Bayesian inference using deep learning, supporting simulation-based inference and multiple backends like JAX, PyTorch, and TensorFlow.

  • Project
    intel/neural-compressor

    An open-source Python library that provides model compression and quantization techniques for PyTorch, TensorFlow, and JAX models to improve inference efficiency.

  • Project
    SciML/DiffEqFlux.jl

    A Julia package that integrates differential equation solvers into neural networks to enable the creation of physically informed scientific machine learning models.